Language Generation on 300K-example corpus (test)
136.02Generation PerplexityDiHAL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DiHALDiffusion target=Hidden state hℓ*, Recovery interface=Retained upper layers + LM head2026.05 | 136.02 | 59.13 | |
| CoDARDiffusion target=Continuous token/latent states, Recovery interface=Continuous-to-discrete recovery2026.05 | 144.83 | 47.77 | |
| LD4LGDiffusion target=Learned text latent, Recovery interface=Frozen BART + latent decoder2026.05 | 166.11 | 57.97 | |
| Diffusion-LMDiffusion target=Token embeddings, Recovery interface=Embedding-to-token recovery2026.05 | 683.43 | 43.24 | |
| SEDDiffusion target=Self-conditioned embeddings, Recovery interface=Embedding-to-token recovery2026.05 | 778.82 | 49.42 |